JP6152786B2 - Communication control apparatus, information processing apparatus, parallel computer system, control program, and parallel computer system control method - Google Patents

Description

  The present invention relates to a communication control device, an information processing device, a parallel computer system, a control program, and a control method for the parallel computer system.

  When executing a parallel application program in a parallel computer system, mainly using a Message Passing Interface (MPI), each of a plurality of processes proceeds in parallel while repeating arithmetic processing and inter-process communication processing. This inter-process communication process is performed between processes in a node of the parallel computer system, and is also performed between processes of different nodes. At this time, the start time of communication processing may differ between processes due to the difference in the computation processing time between processes.

  For example, as shown in FIG. 1, when the process P0, the process P1, and the process P2 execute the parallel application program, the process P0 and the process P1 that have completed the arithmetic processing earlier try communication processing with the process P2, respectively. . However, since the process P2 is performing arithmetic processing and communication processing cannot be started, the process P0 and the process P1 each wait for completion of the arithmetic processing by the process P2. During this time, since the process P0 and the process P1 do not perform arithmetic processing or communication processing, the utilization efficiency of the parallel computer system is lowered and the number effect is deteriorated.

  In order to solve this problem, a method in which a developer of a parallel application program performs code rewriting, parameter tuning, etc. to equalize the processing time of each process and aligns the start time of communication processing among the processes.

  As a method for confirming whether or not the start time of communication processing is uniform among processes, there is known a method for confirming based on a value called a synchronization waiting time. The synchronization waiting time is obtained, for example, as follows.

1. The start time is acquired for each communication process of each process. This start time can be acquired, for example, as an elapsed time from the execution start time of each process.
2. The maximum value of the start times of a plurality of processes is obtained for each communication process.
3. The difference between the maximum value and the communication processing start time of each process is obtained, and the differences regarding a plurality of communication processes are integrated for each process, and recorded as the synchronization waiting time.

  The synchronization waiting time of each process becomes 0 when the start times of all communication processes coincide between processes, and the greater the difference in start time between processes, the closer to the elapsed time required to execute the parallel application program. Therefore, it can be determined that the closer the synchronization waiting time is to 0, the more desirable the state is.

  In the first communication process of FIG. 1, the start times of the process P0, the process P1, and the process P2 are 20, 10, and 30, respectively, and the maximum start time is 30 indicated by the arrow 101. Differences between the maximum value 30 and the start times of the process P0, the process P1, and the process P2 are 10, 20, and 0, respectively.

  In the second communication process, the start times of the process P0, the process P1, and the process P2 are 60, 70, and 50, respectively, and the maximum start time is 70 indicated by the arrow 102. Differences between the maximum value 70 and the start times of the process P0, the process P1, and the process P2 are 10, 0, and 20, respectively.

  Therefore, when the differences related to the first and second communication processes are integrated, the synchronization waiting times of the process P0, the process P1, and the process P2 are 20, 20, and 20, respectively. In this case, the elapsed time required to execute the parallel application program is 80, and 20 of them can be interpreted as wasted time for waiting for completion of the arithmetic processing by another process.

  The parallel application performance profiling tool of Cray Corporation in the United States hooks the MPI collective communication function to obtain the synchronization waiting time, and automatically calls the inter-process synchronization interface (MPI_Barrier function) before the start of communication processing. Then, the parallel application performance profiling tool obtains the total elapsed time of the MPI_Barrier function for each process.

  There is also known a reduction operation device that performs a reduction operation for obtaining the sum, maximum value, minimum value, etc. of data for data of a plurality of processes (see, for example, Patent Document 1).

JP 2010-122848 A

  In the following description, the parallel application program may be simply referred to as a parallel application.

The conventional parallel computer system has the following problems.
When the inter-process synchronization interface is called to obtain the synchronization waiting time as in the parallel application performance profiling tool, the synchronization processing is added to the processing of each process, so the behavior of the parallel application may be greatly changed. For this reason, it is desirable to obtain the maximum value of the start time for each communication process without performing the synchronization process.

  To obtain the maximum start time for each communication process without performing synchronous processing (asynchronously), each process acquires the start time for each communication process and writes it to the main storage area before executing the parallel application. A method of calculating the maximum value during or after execution is conceivable.

  When the maximum value is calculated after the execution of the parallel application, data of the start time of all communication processes is recorded. Therefore, in a parallel application having a long execution time, the main storage area that can be used by the parallel application may be compressed. In such a case, it is desirable to calculate the maximum value during execution of the parallel application.

  When calculating the maximum value during the execution of the parallel application, it is not necessary to record the data of the start time of all the communication processes, and it is sufficient to record the integrated value of the differences regarding the past communication processes as the synchronization waiting time. Therefore, even in a parallel application having a long execution time, the main storage area is not compressed. In this case, a method of transmitting and receiving the communication processing start time of each process using an asynchronous communication interface of software or hardware during the arithmetic processing is conceivable.

  However, in the method using the asynchronous communication interface of software, the calculation of the maximum value is software processing by the Central Processing Unit (CPU), so that the behavior of the parallel application may change by consuming CPU time.

  In the method using the hardware asynchronous communication interface, each process can transfer start time data to a network reduction mechanism mounted on the network interface and instruct maximum value calculation.

  Here, when the calculation processing time between one communication process and the next communication process is very short, there is a possibility that the maximum value calculation by the network reduction mechanism may not be completed. If the network reduction mechanism can execute only a single maximum value operation at the same time, each process cannot instruct the subsequent maximum value operation until the preceding maximum value operation is completed, and the behavior of the parallel application is changed.

  In one aspect, an object of the present invention is to perform an operation using the start time of interprocess communication without changing the behavior of a parallel application during execution of the parallel application.

  In one proposal, the communication control device connected to the arithmetic processing device and the main storage device includes an arithmetic unit.

  The first sequence information is given by the arithmetic processing device to the first start time at which the first process of the plurality of processes included in the program executed by the arithmetic processing device starts the first inter-process communication. Sequence information written to the main memory. When the second sequence information given at the second start time at which the second process of the plurality of processes has started the second inter-process communication is newer than the first sequence information, The calculation using the first start time is not performed.

  When the second sequence information corresponds to the first sequence information, the calculation unit performs a calculation using the first start time and the second start time and outputs a calculation result.

  According to the embodiment, it is possible to perform an operation using the start time of interprocess communication without changing the behavior of the parallel application during execution of the parallel application.

It is a figure which shows the inter-process communication process in a parallel computer system. It is a figure which shows the inter-process communication process at the time of using an inter-process synchronous interface. It is a figure which shows a network reduction mechanism. It is a block diagram of a 1st communication control apparatus. It is a block diagram of a parallel computer system. It is a block diagram of a node. It is a block diagram of a 2nd communication control apparatus. It is a flowchart of the arithmetic processing using the start time of interprocess communication. It is a flowchart which shows the 1st specific example of a calculation process. It is a flowchart which shows the 2nd specific example of a calculation process.

Hereinafter, embodiments will be described in detail with reference to the drawings.
When the inter-process synchronization interface is called to obtain the synchronization waiting time as in the parallel application performance profiling tool, the synchronization processing is added to the processing of each process, so the behavior of the parallel application may be greatly changed.

  FIG. 2 shows an example in which the inter-process synchronization interface is called by the parallel application performance profiling tool in the parallel application of FIG. In this case, since the synchronization process for obtaining the synchronization waiting time is performed after the communication process of each process is completed, the elapsed time required to execute the parallel application is increased by 25% from 80 to 100. Although it depends on the parallel application, there are many cases where the elapsed time increases by 10% or more as shown in FIG. For this reason, it is desirable to obtain the maximum value of the start time for each communication process without performing the synchronization process.

  To obtain the maximum start time for each communication process without performing synchronous processing (asynchronously), each process acquires the start time for each communication process and writes it to the main storage area before executing the parallel application. A method of calculating the maximum value during or after execution is conceivable.

  First, when the maximum value is calculated after execution of a parallel application, there is a possibility that a parallel application having a long execution time may squeeze the main storage area that the parallel application can use. For example, it is assumed that communication processing occurs at a frequency of once per 1 ms in a parallel application in which 100,000 processes run for one day. If 8 bytes of start time data is recorded for each communication process, about 691 Mbytes of main storage area per process is filled with start time data just before the end of the parallel application.

  In a parallel computer system equipped with a main memory device of 32 GB per node, when 16 processes operate in parallel in each node, the main memory area of about 11 GB is filled with start time data to execute parallel applications. There is a possibility of trouble. In such a case, it is desirable to calculate the maximum value during execution of the parallel application.

  When calculating the maximum value during the execution of the parallel application, it is not necessary to record the data of the start time of all the communication processes, and it is sufficient to record the integrated value of the differences regarding the past communication processes as the synchronization waiting time. Therefore, even in a parallel application having a long execution time, the main storage area is not compressed. In this case, a method of transmitting and receiving the communication processing start time of each process using an asynchronous communication interface of software or hardware during the arithmetic processing is conceivable.

  In the method using the asynchronous communication interface of software, the data of the start time can be transmitted and received between processes by the MPI non-blocking communication interface (MPI_Isend function, MPI_Irecv function, MPI_Wait function, etc.). However, since the calculation of the maximum value is software processing by the Central Processing Unit (CPU), the behavior of the parallel application may change by consuming CPU time.

  In the method using the hardware asynchronous communication interface, each process can transfer start time data to a network reduction mechanism mounted on the network interface and instruct maximum value calculation. Network reduction mechanism is hardware that receives data from each process of each node connected in the communication network, performs a single operation on all data, and holds the operation result in each node Refers to mechanism. Examples of the calculation by the network reduction mechanism include calculation for obtaining the sum, maximum value, minimum value, and the like of data.

  For example, each process reads the maximum value from the network reduction mechanism at the start of the next communication process after transferring the communication process start time to the network reduction mechanism at the completion of a certain communication process and instructing the maximum value calculation. Can do. Thereby, the network reduction mechanism can calculate the maximum value during the arithmetic processing of each process.

  FIG. 3 shows an example in which the network reduction mechanism 301 calculates the maximum value in the parallel application of FIG. In this case, the network reduction mechanism 301 calculates the maximum value 30 of the start time of the first communication process during the second calculation process. Then, the process P0, the process P1, and the process P2 set the differences 10, 20, and 0 between the maximum value 30 and the start times 20, 10, and 30 of the first communication process at the start of the second communication process. Ask for each.

  Here, when the calculation processing time between one communication process and the next communication process is very short, there is a possibility that the maximum value calculation by the network reduction mechanism may not be completed. If the network reduction mechanism can execute only a single maximum value operation at the same time, each process cannot instruct the subsequent maximum value operation until the preceding maximum value operation is completed, and the behavior of the parallel application is changed.

  There are several ways to prevent the behavior of parallel applications from changing when using a network reduction mechanism. For example, there is a method in which the network reduction mechanism skips the instruction for the subsequent maximum value calculation until the preceding maximum value calculation is completed. However, since the calculation by the network reduction mechanism is not always completed between processes at the same time, a situation may occur in which the number of skips differs for each process. When the subsequent maximum value calculation is performed in such a situation, the maximum value of the start times of a plurality of different communication processes is calculated instead of the maximum value of the start time of one communication process.

  As another method, a plurality of network reduction mechanisms are implemented in the network interface, and the network reduction mechanisms are used in order. However, the increase in hardware becomes very large. Since it is a rare problem that the maximum value calculation by the network reduction mechanism is not completed, it is not realistic to apply a large hardware cost to solve the problem.

  Such a problem occurs not only when calculating the maximum value of the time when interprocess communication is started in a parallel application, but also when performing other operations using the time when communication between processes is started.

  FIG. 4 shows a configuration example of the communication control apparatus of the embodiment. The communication control device 401 in FIG. 4 is provided in an information processing device (computer) corresponding to each node of the parallel computer system, and includes a calculation unit 411.

  The arithmetic processing unit in the information processing apparatus executes a program, and a first sequence with respect to a first start time when a first process among the plurality of processes included in the program starts communication between the first processes. Give information. Then, the first start time and the first sequence information are written to the main storage device in the information processing apparatus.

  When the second sequence information given at the second start time when the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, the communication control device 401 The calculation unit 411 does not perform calculation using the first start time. On the other hand, when the second sequence information corresponds to the first sequence information, the calculation unit 411 performs a calculation using the first start time and the second start time and outputs a calculation result.

  The sequence information is information indicating the order of communication processing performed by each process, and the same sequence information is given to the communication processing in the same order among a plurality of processes. For example, the sequence information given to the first communication process performed by the second process is the same as the sequence information given to the first communication process performed by the first process. Further, the sequence information given to the second communication process performed by the second process is the same as the sequence information given to the second communication process performed by the first process.

  By comparing the sequence information given at the respective start times between the first and second processes, it is possible to determine whether the two start times are the start times of the communication processes in the same order. it can.

  Then, the calculation unit 411 uses the first start time when the second sequence information given at the second start time is newer than the first sequence information given at the first start time. Do not perform the operation. For this reason, the first process can sequentially instruct the calculation unit 411 to perform the calculation using the latest start time without waiting for the completion of the calculation using the old start time. Therefore, during the execution of the parallel application, it is possible to perform an operation using the start time of the interprocess communication without changing the behavior of the parallel application.

  FIG. 5 shows a configuration example of the parallel computer system of the embodiment. The parallel computer system 500 of FIG. 5 includes n (n is an integer of 1 or more) computation nodes and disk nodes 502, which are nodes 501-1 to 501-n. The nodes 501-1 to 501-n and the disk node 502 are connected to each other by a communication network 503.

  FIG. 6 shows a configuration example of an information processing apparatus corresponding to the nodes 501-i (i = 1 to n) in FIG. 5. A node 501-i in FIG. 6 includes a CPU 601, a memory 602, a medium driving device 603, and a communication control device 401. The CPU 601, the memory 602, the medium driving device 603, and the communication control device 401 are connected to each other via a bus 604.

The memory 602 corresponds to the main storage device. The memory 602 is a semiconductor memory such as a read only memory (ROM) or a random access memory (RAM), and stores a parallel application program and data used for the processing.

  The CPU 601 corresponds to an arithmetic processing unit (processor) and can include a memory controller and a network interface controller. For example, the CPU 601 uses the memory 602 to execute the parallel application program. The CPU 601 can also execute programs such as an operating system (OS) and a network interface driver.

  The medium driving device 603 drives the portable recording medium 605 and accesses the recorded contents. The portable recording medium 605 is a memory device, a flexible disk, an optical disk, a magneto-optical disk, or the like. The portable recording medium 605 may be a compact disk read only memory (CD-ROM), a digital versatile disk (DVD), a universal serial bus (USB) memory, or the like. A user or an operator can store parallel application programs and data in the portable recording medium 605 and load them into the memory 602 for use.

  The communication control device 401 is a network interface that is connected to the communication network 503 in FIG. 5 and communicates with other nodes. The node 501-i can receive the parallel application program and data from a device external to the parallel computer system via the communication control device 401 and load them into the memory 602 for use.

  Note that the node 501-i does not have to include all the components shown in FIG. 6, and some of the components can be omitted depending on applications and conditions. For example, when the portable recording medium 605 is not used, the medium driving device 603 may be omitted.

  The disk node 502 in FIG. 5 has a configuration in which a disk device is added to the configuration of the node 501-i in FIG. The disk device is, for example, a magnetic disk device, an optical disk device, a magneto-optical disk device, or the like, and may be a hard disk drive. The disk device can store data used for processing parallel application programs.

  For example, the nodes 501-1 to 501-n execute parallel applications using data stored in the disk node 502. At this time, one or more processes are generated in each node, and a plurality of processes in the nodes 501-1 to 501-n are parallelized by MPI to perform collective communication for all processes. repeat.

  If a disk device is provided for each of the nodes 501-1 to 501-n, the disk node 502 may be omitted.

  FIG. 7 shows a configuration example of the calculation unit 411 in the communication control apparatus 401 of FIG. The arithmetic unit 411 in FIG. 7 includes a CPU 701, a memory 702, an interface 703, and an interface 704. The CPU 701, the memory 702, the interface 703, and the interface 704 are connected to each other via a bus 705.

  The CPU 701 corresponds to an arithmetic processing device (processor), and the memory 702 corresponds to a main storage device. The memory 702 is a semiconductor memory such as a ROM and a RAM, for example, and stores a control program and data. Data stored in the memory 702 includes data transferred from the CPU 601 in FIG. 6, data received from other nodes, and the like. For example, the CPU 701 executes a control program using the memory 702 to perform an operation of the network reduction mechanism, and writes the operation result in the memory 702.

  The user or operator can store the control program and data in the portable recording medium 605 of FIG. 6 and load them into the memory 702 for use. As described above, the computer-readable recording medium that stores the control program and data is a physical (non-transitory) recording medium such as the memory 702 or the portable recording medium 605.

  The interface 703 is connected to the bus 604 in FIG. 6 to communicate with the CPU 601, and the interface 704 is connected to the communication network 503 in FIG. 5 to communicate with other nodes. The calculation is made efficient by transmitting / receiving partial calculation results of the network reduction mechanism between the nodes via the interface 704. The arithmetic unit 411 can receive a control program and data from a device external to the parallel computer system via the interface 703 or the interface 704 and load them into the memory 702 for use.

  Note that the configuration of the calculation unit 411 is not limited to the configuration of FIG. 7, and part or all of the processing can be implemented by wiring logic.

  FIG. 8 is a flowchart illustrating an example of arithmetic processing using the start time of interprocess communication performed in the node 501-1. In this example, the first process is a process generated in the node 501-1 among a plurality of processes of the parallel application. The second process is a process generated in the node 501-1 among a plurality of processes of the parallel application, or a process generated in any of the nodes 501-2 to 501-n.

  The CPU 601 in the node 501-1 gives the first sequence information to the first start time when the first process starts the first inter-process communication. Then, the CPU 601 writes the first start time and the first sequence information to the memory 602 in the node 501-1, and writes the first start time and the first sequence information in the memory 702 in the arithmetic unit 411. Forward to. The communication control device 401 in the node 501-1 transmits the first start time and the first sequence information to the node 501-2 to the node 501-n.

  When the second process is generated in the node 501-1, the CPU 601 in the node 501-1 sets the second process time for the second start time when the second process starts the second inter-process communication. Assign sequence information. Then, the CPU 601 writes the second start time and the second sequence information in the memory 602 in the node 501-1, and writes the second start time and the second sequence information in the memory 702 in the arithmetic unit 411. Forward to. The communication control device 401 in the node 501-1 transmits the second start time and the second sequence information to the node 501-2 to the node 501-n.

  When the second process is generated in the node 501-2, the CPU 601 in the node 501-2 sets the second process time to the second start time when the second process starts the second inter-process communication. Assign sequence information. Then, the CPU 601 writes the second start time and the second sequence information to the memory 602 in the node 501-2, and writes the second start time and the second sequence information in the memory 702 in the arithmetic unit 411. Forward to. The communication control device 401 in the node 501-2 transmits the second start time and the second sequence information to the node 501-1 and the nodes 501-3 to 501-n.

  Similarly, each CPU 601 of the node 501-3 to the node 501-n gives sequence information to the start time when each process starts inter-process communication. Then, the CPU 601 writes the start time and sequence information to the memory 602 and transfers the start time and sequence information to the memory 702 in the calculation unit 411. The communication control devices 401 of the nodes 501-3 to 501-n transmit the start time and sequence information to other nodes.

  In node 501-1, CPU 701 in operation unit 411 checks whether or not the second sequence information given at the second start time is newer than the first sequence information (step 801).

  When the second sequence information is newer than the first sequence information (step 801, YES), the CPU 701 does not perform calculation using the first start time. On the other hand, when the check result in step 801 is NO, the CPU 701 checks whether or not the second sequence information corresponds to the first sequence information (step 802).

  When the second sequence information corresponds to the first sequence information (step 802, YES), the CPU 701 performs calculation using at least the first start time and the second start time, and stores the calculation result in the memory. Write to 702 (step 803).

  On the other hand, if the check result in step 802 is NO, the CPU 701 does not perform the calculation using the second start time, and the third start time and the third start time when the second process starts the third inter-process communication. Until the sequence information is received (step 804). When the third sequence information corresponds to the first sequence information, the CPU 701 performs a calculation using at least the first start time and the third start time, and writes the calculation result in the memory 702.

  The calculation units 411 of the nodes 501-2 to 501-n can perform the same calculation process as the calculation unit 411 of the node 501-1.

  FIG. 9 is a flowchart illustrating a specific example of the arithmetic processing in FIG. 8 performed when one process is operating in each of the nodes 501-1 to 501-n in FIG.

  As the sequence information, for example, a sequence number that is a non-negative integer in ascending order is used. The sequence number is managed by the network interface driver and is reset to 0 at the start of the parallel application. The memory 602 in each node 501-i (i = 1 to n) stores a start time T (i) at which a process starts inter-process communication and a sequence number S (i) of the process.

  The CPU 601 in each node 501-i operates a process by executing a parallel application, and operates a network interface driver by executing a network interface driver program.

  When a process in the node 501-i starts inter-process communication, the process instructs the network interface driver to perform arithmetic processing (step 901). Next, the CPU 601 operates as a network interface driver, and reads the start time T (i) and the sequence number S (i) from the memory 602 (step 902). Next, the CPU 601 writes the read start time T (i) and sequence number S (i) in the memory 702 in the calculation unit 411 and instructs the CPU 701 in the calculation unit 411 to perform calculation processing. Then, the CPU 601 increments the sequence number S (i) in the memory 602 by 1.

  Next, the CPU 701 transmits the start time T (i) and the sequence number S (i) written in the memory 702 to another node 501-j (1 ≦ j ≦ n, j ≠ i) (step 903). ). Then, the CPU 701 checks whether or not the start times and sequence numbers of all processes of the parallel application are written in the memory 702 (step 904).

  When the start time and sequence number of any process are not written (step 904, NO), the CPU 701 checks whether the start time and sequence number have been received from another node (step 908). .

  When the start time and the sequence number have not been received from another node (step 908, NO), the CPU 701 repeats the processing after step 904. On the other hand, when the start time T (k) and the sequence number S (k) are received from another node 501-k (1 ≦ k ≦ n, k ≠ i) (step 908, YES), the CPU 701 has received The sequence number S (k) is compared with the sequence number S (i) (step 909).

  When the received sequence number S (k) is smaller than the sequence number S (i) (step 909, NO), the CPU 701 repeats the processing after step 904. On the other hand, when the received sequence number S (k) is greater than or equal to the sequence number S (i) (step 909, YES), the CPU 701 stores the received start time T (k) and the sequence number S (k) in the memory 702. (Step 910). The CPU 701 again compares the received sequence number S (k) with the sequence number S (i) (step 911).

  When the received sequence number S (k) matches the sequence number S (i) (step 911, NO), the CPU 701 repeats the processing after step 904. Thus, when the start time and sequence number of all processes are written in the memory 702 (step 904, YES), the CPU 701 obtains the maximum value of the start time of all processes and writes it in the memory 702 (step 905).

  Next, the CPU 701 interrupts the CPU 601 (step 906), and the CPU 601 operates the network interface driver. The CPU 601 reads out the maximum value of the start time from the memory 702, writes it in the memory 602, and initializes the memory 702 (step 907). As a result, the start times and sequence numbers of all processes in the memory 702 are deleted.

  On the other hand, when the received sequence number S (k) is larger than the sequence number S (i) (step 911, YES), the CPU 701 interrupts the CPU 601 and notifies that the instructed arithmetic processing is stopped (step 912). . By this notification, the CPU 601 stops the arithmetic processing instructed by the process to the network interface driver without initializing the memory 702. Thereafter, when the process instructs the network interface driver to perform the next calculation process, the CPU 601 starts the process from step 901 onward.

  FIG. 10 is a flowchart illustrating a specific example of the arithmetic processing in FIG. 8 performed when a plurality of processes are operating in each of the nodes 501-1 to 501-n in FIG.

  As the sequence information, for example, a sequence number that is a non-negative integer in ascending order is used. The sequence number is managed by the network interface driver and is reset to 0 at the start of the parallel application. Assume that the total number of processes operating in the nodes 501-1 to 501-n is p, and the identification information of the mth process is P (m) (m = 1 to p). The memory 602 in each node 501-i (i = 1 to n) includes a start time T (m) at which each process in the node 501-i starts inter-process communication, and a sequence number S (m) of the process. And remember.

  When the process P (r) among the plurality of processes in the node 501-i starts inter-process communication, the process P (r) instructs the network interface driver to perform arithmetic processing (step 1001). Next, the CPU 601 operates as a network interface driver, and reads the start time T (r) and the sequence number S (r) of the process P (r) from the memory 602 (step 1002). Next, the CPU 601 writes the read start time T (r) and sequence number S (r) to the memory 702 in the calculation unit 411 and instructs the CPU 701 in the calculation unit 411 to perform calculation processing. Then, the CPU 601 increments the sequence number S (r) in the memory 602 by 1.

  Next, the CPU 701 transmits the start time T (r) and the sequence number S (r) written in the memory 702 to another node 501-j (1 ≦ j ≦ n, j ≠ i) (step 1003). ).

  Thereafter, when another process P (x) in the node 501-i starts inter-process communication, the process P (x) instructs the network interface driver to perform arithmetic processing, similarly to step 1001. Similarly to step 1002, the CPU 601 writes the start time T (x) of the process P (x) and the sequence number S (x) in the memory 702, and sets the sequence number S (x) in the memory 602 to only 1. Increment. Then, as in step 1003, the CPU 701 transfers the start time T (x) and the sequence number S (x) written in the memory 702 to another node 501-j (1 ≦ j ≦ n, j ≠ i). Send.

  Next, the CPU 701 checks whether or not the start time and sequence number of all processes of the parallel application are written in the memory 702 (step 1004). When the start time and sequence number of any process are not written (step 1004, NO), the CPU 701 checks whether the start time and sequence number are received from another node (step 1008). .

  When the start time and the sequence number have not been received from another node (step 1008, NO), the CPU 701 repeats the processing after step 1004.

  During this time, when another process P (y) in the node 501-i starts inter-process communication, the process P (y) instructs the network interface driver to perform arithmetic processing in the same manner as in Step 1001. Similarly to step 1002, the CPU 601 writes the start time T (y) of the process P (y) and the sequence number S (y) to the memory 702, and sets the sequence number S (y) in the memory 602 to only 1. Increment. Then, as in step 1003, the CPU 701 transfers the start time T (y) and the sequence number S (y) written in the memory 702 to another node 501-j (1 ≦ j ≦ n, j ≠ i). Send.

  On the other hand, when the start time T (q) and the sequence number S (q) are received from another node 501-k (1 ≦ k ≦ n, k ≠ i) (step 1008, YES), the CPU 701 proceeds to step 1009. Perform the process.

  In step 1009, the CPU 701 obtains the maximum value of the sequence numbers corresponding to one or more processes in the node 501-i among the sequence numbers written in the memory 702. Then, the CPU 701 compares the received sequence number S (q) with the maximum value of the sequence number. When the received sequence number S (q) is smaller than the maximum value of the sequence number (step 1009, NO), the CPU 701 repeats the processing after step 1004.

  On the other hand, when the received sequence number S (q) is greater than or equal to the maximum value of the sequence number (step 1009, YES), the CPU 701 stores the received start time T (q) and the sequence number S (q) in the memory 702. Write (step 1010). The CPU 701 again compares the received sequence number S (q) with the maximum sequence number (step 1011).

  When the received sequence number S (q) matches the maximum value of the sequence number (step 1011, NO), the CPU 701 repeats the processing after step 1004. When the start times and sequence numbers of all processes in the node 501-i are written in the memory 702 and these sequence numbers match the sequence number S (r), the maximum value of the sequence number obtained in step 1009 is S ( r).

  When the start times and sequence numbers of all processes of the parallel application are written in the memory 702 (step 1004, YES), the CPU 701 obtains the maximum start time of all processes and writes it in the memory 702 (step 1005).

  Next, the CPU 701 interrupts the CPU 601 (step 1006), and the CPU 601 operates the network interface driver. Then, the CPU 601 reads the maximum value of the start time from the memory 702, writes it in the memory 602, and initializes the memory 702 (step 1007). As a result, the start times and sequence numbers of all processes in the memory 702 are deleted.

  On the other hand, when the received sequence number S (q) is larger than the maximum value of the sequence number (step 1011, YES), the CPU 701 interrupts the CPU 601 and notifies it to stop the instructed arithmetic processing (step 1012). With this notification, the CPU 601 stops the arithmetic processing instructed by the process P (r) to the network interface driver without initializing the memory 702. Thereafter, when the process P (r) or another process instructs the network interface driver to perform the next arithmetic processing, the CPU 601 starts the processing from step 1001 onward.

  According to the arithmetic processing in FIG. 9 or 10, the arithmetic unit 411 uses the start time until the start times of all processes having the latest sequence number are written in the memory 702 even if the arithmetic processing is instructed. Do not do. Then, when the sequence number of the process in the node 501-i is incremented to be the same as the received sequence number, arithmetic processing using the start times of all processes having that sequence number is performed.

  In this way, calculation processing is always performed for the start time having the latest sequence number, and each process uses the latest start time without waiting for completion of the calculation processing using the old start time. Can be instructed to the calculation unit 411 one after another. Therefore, during the execution of the parallel application, it is possible to perform an operation using the start time of the interprocess communication without changing the behavior of the parallel application.

  In addition, it is only necessary to provide one arithmetic unit 411 in the communication control device 401, and it is not necessary to provide a plurality of network reduction mechanisms, so that an increase in hardware can be suppressed. In particular, in a large-scale parallel computer system, a great effect can be expected from the arithmetic processing of FIG. 9 or FIG.

  The configurations of the communication control device 401 in FIGS. 4 and 7, the parallel computer system 500 in FIG. 5, and the node 501-i in FIG. 6 are merely examples, and some of the components depend on the use and conditions of the parallel computer system May be omitted or changed.

  The flowcharts of FIGS. 8 to 10 are merely examples, and some processes may be omitted or changed depending on the configuration and conditions of the parallel computer system. For example, in the arithmetic processing of FIG. 9 or FIG. 10, information indicating the order of communication processing other than the sequence number can be used as sequence information. Further, in step 905 of FIG. 9 or step 1005 of FIG. 10, the CPU 701 can obtain other numerical values such as the sum of the start times and the minimum value of the start times instead of obtaining the maximum value of the start times.

  Although the disclosed embodiments and their advantages have been described in detail, those skilled in the art can make various modifications, additions and omissions without departing from the scope of the present invention as explicitly set forth in the claims. Let's go.

Regarding the embodiment described with reference to FIGS. 4 to 10, the following additional notes are disclosed.
(Appendix 1)
A communication control device connected to the arithmetic processing device and the main storage device,
The main processing unit assigns the main memory to a first start time at which a first process among the plurality of processes included in a program executed by the arithmetic processing unit starts a first inter-process communication. The second sequence information given at the second start time at which the second process of the plurality of processes started the second inter-process communication is newer than the first sequence information written in the apparatus. In the case where the calculation using the first start time is not performed and the second sequence information corresponds to the first sequence information, the first start time and the second start time are used. A computation unit that performs the computation and outputs the computation result,
A communication control device comprising:
(Appendix 2)
The first sequence information is newer than the second sequence information, and the third sequence information given at the third start time when the second process starts the third inter-process communication is In the case of corresponding to the first sequence information, the calculation unit does not perform calculation using the second start time, but performs calculation using the first start time and the third start time. The communication control apparatus according to appendix 1, wherein a calculation result is output.
(Appendix 3)
The arithmetic processing unit executes the first process, and the arithmetic processing unit included in the information processing apparatus connected via the communication control unit executes the second process. Or the communication control apparatus of 2.
(Appendix 4)
The first sequence information includes a fourth start time when a fourth process executed by the arithmetic processing unit that executes the first process among the plurality of processes starts a fourth inter-process communication. The communication control device according to supplementary note 3, wherein the communication control device corresponds to the fourth sequence information given to or is newer than the fourth sequence information.
(Appendix 5)
The communication control apparatus according to appendix 1 or 2, wherein the arithmetic processing unit executes the first process and the second process.
(Appendix 6)
An arithmetic processing unit that executes a program and assigns first sequence information to a first start time at which a first process among the plurality of processes included in the program starts first inter-process communication;
A main storage device for storing the first start time and the first sequence information;
When the second sequence information given at the second start time at which the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, When the calculation using the start time of 1 is not performed and the second sequence information corresponds to the first sequence information, the calculation using the first start time and the second start time is performed. A communication control device for outputting a calculation result,
An information processing apparatus comprising:
(Appendix 7)
A parallel computer system having a plurality of information processing devices,
At least one information processing apparatus among the plurality of information processing apparatuses is
An arithmetic processing unit that executes a program and assigns first sequence information to a first start time at which a first process among the plurality of processes included in the program starts first inter-process communication;
A main storage device for storing the first start time and the first sequence information;
When the second sequence information given at the second start time at which the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, When the calculation using the start time of 1 is not performed and the second sequence information corresponds to the first sequence information, the calculation using the first start time and the second start time is performed. A communication control device for outputting a calculation result,
A parallel computer system characterized by comprising:
(Appendix 8)
A control program for an information processing device having an arithmetic processing device, a communication control device, and a main storage device,
The main processing unit assigns the main memory to a first start time at which a first process among the plurality of processes included in a program executed by the arithmetic processing unit starts a first inter-process communication. The second sequence information given at the second start time at which the second process of the plurality of processes started the second inter-process communication is newer than the first sequence information written in the apparatus. In the case where the calculation using the first start time is not performed and the second sequence information corresponds to the first sequence information, the first start time and the second start time are used. Output the calculation result.
A control program that causes an arithmetic processing unit in the communication control unit to execute processing.
(Appendix 9)
A method for controlling a parallel computer system having a plurality of information processing devices,
At least one information processing device of the plurality of information processing devices is
Run the program
The first sequence information is given to the first start time when the first process among the plurality of processes included in the program starts the first inter-process communication,
When the second sequence information given at the second start time at which the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, When the calculation using the start time of 1 is not performed and the second sequence information corresponds to the first sequence information, the calculation using the first start time and the second start time is performed. Output the calculation result.
A control method characterized by that.

101, 102 arrow 301 network reduction mechanism 401 communication control device 411 arithmetic unit 500 parallel computer system 501-1 to 501-n, 501-i node 502 disk node 503 communication network 601, 701 CPU
602, 702 Memory 603 Medium drive device 604, 705 Bus 605 Portable recording medium 703, 704 Interface

Claims (7)

A communication control device connected to the arithmetic processing device and the main storage device,
The main processing unit assigns the main memory to a first start time at which a first process among the plurality of processes included in a program executed by the arithmetic processing unit starts a first inter-process communication. The second sequence information given at the second start time at which the second process of the plurality of processes started the second inter-process communication is newer than the first sequence information written in the apparatus. In the case where the calculation using the first start time is not performed and the second sequence information corresponds to the first sequence information, the first start time and the second start time are used. A computation unit that performs the computation and outputs the computation result,
A communication control device comprising:   The first sequence information is newer than the second sequence information, and the third sequence information given at the third start time when the second process starts the third inter-process communication is In the case of corresponding to the first sequence information, the calculation unit does not perform calculation using the second start time, but performs calculation using the first start time and the third start time. The communication control apparatus according to claim 1, wherein a calculation result is output.   The arithmetic processing device executes the first process, and the arithmetic processing device included in an information processing device connected via the communication control device executes the second process. The communication control apparatus according to 1 or 2. An arithmetic processing unit that executes a program and assigns first sequence information to a first start time at which a first process among the plurality of processes included in the program starts first inter-process communication;
A main storage device for storing the first start time and the first sequence information;
When the second sequence information given at the second start time at which the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, When the calculation using the start time of 1 is not performed and the second sequence information corresponds to the first sequence information, the calculation using the first start time and the second start time is performed. A communication control device for outputting a calculation result,
An information processing apparatus comprising: A parallel computer system having a plurality of information processing devices,
At least one information processing apparatus among the plurality of information processing apparatuses is
An arithmetic processing unit that executes a program and assigns first sequence information to a first start time at which a first process among the plurality of processes included in the program starts first inter-process communication;
A main storage device for storing the first start time and the first sequence information;
When the second sequence information given at the second start time at which the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, When the calculation using the start time of 1 is not performed and the second sequence information corresponds to the first sequence information, the calculation using the first start time and the second start time is performed. A communication control device for outputting a calculation result,
A parallel computer system characterized by comprising: A control program for an information processing device having an arithmetic processing device, a communication control device, and a main storage device,
The main processing unit assigns the main memory to a first start time at which a first process among the plurality of processes included in a program executed by the arithmetic processing unit starts a first inter-process communication. The second sequence information given at the second start time at which the second process of the plurality of processes started the second inter-process communication is newer than the first sequence information written in the apparatus. In the case where the calculation using the first start time is not performed and the second sequence information corresponds to the first sequence information, the first start time and the second start time are used. Output the calculation result.
A control program that causes an arithmetic processing unit in the communication control unit to execute processing. A method for controlling a parallel computer system having a plurality of information processing devices,
At least one information processing device of the plurality of information processing devices is
Run the program
The first sequence information is given to the first start time when the first process among the plurality of processes included in the program starts the first inter-process communication,
When the second sequence information given at the second start time at which the second process of the plurality of processes starts the second inter-process communication is newer than the first sequence information, When the calculation using the start time of 1 is not performed and the second sequence information corresponds to the first sequence information, the calculation using the first start time and the second start time is performed. Output the calculation result.
A control method characterized by that.

Images